Containing the Gear: Deep Dive on SELinux, Multi-tenancy, Containers & Security with Dan Walsh
Presenter: Dan Walsh
In this talk, Dan will do a deep dive into the Origin PaaS use of SELinux and containerization. He will discuss how SELinux being utilized to ensure that Origin is the the most secure PAAS available today. He will address some of his ideas for the future of Origin and SELinux.
From 2013-04-14 OpenShift Origin Community Day in Portland, Oregon
2. SELinux is a LABELING System
● Everything has a label
– Process,file,dir, chr_file, blk_file, port, node.
● SELinux Policy defines that access between
process labels and all other labels.
● The Kernel controls the access.
5. SELinux is Type Enforcement
● Process Labels can be on Files
● File Labels can not on Processes
● openshift_t -> Process
● openshift_var_lib_t -> File
6. SELinux is MCS
● system_u:system_r:openshift_t:s0:c1,c2
● Multi Category System
● MCS Separation is for like types, but totally separated
● openshift_t:s0:c1,c2 -> openshift_var_lib_t:s0:c1,c2
● openshift_t:s0:c3,c4 -> openshift_var_lib_t:s0:c3,c4
7. Libvirt – Dynamic Labeling in action
openshift_t:MCS1
openshift_t:MCS2
Kernel
SELinux
Host Hardware
memory, storage, etc.
openshift_t:MCS1 openshift_t:MCS2
8. MCS Labeling based on UID
def gen_level(uid):
SETSIZE=1023
TIER=SETSIZE
ORD=uid;
while ORD > TIER:
ORD = ORD - TIER;
TIER= TIER - 1;
TIER = SETSIZE - TIER;
ORD = ORD + TIER;
return "s0:c%d,c%d" % (TIER, ORD)
9. How do the labels get on gears
● Host receives packet for a gear
– OpenShift server
● launches application with correct SELinux label.
● Sends packet to application
● If connection comes in via git or ssh
– Ssh uses pam_openshift
● Launch sh with correct context
● Launch git with correct context
12. Problems with OpenShift Security
● Gear Application == Administrator of Gear
– Same UID
– Same SELinux Label openshift_t
● Solution:
– openshift_t
● Administrator of gear
– openshift_app_t
● Type of the application
– openshift_var_lib_t
● openshift_t can read/write/execute
● openshift_app_t can read/execute
– openshift_rw_file_t
● openshift_t & openshift_app_t can read/write/execute
13. Problem with OpenShift Security
● All gears run as openhift_t
– All have same network access.
● openshift_t/openshift_app_t
● openshift_net_t/openshift_net -app_t
● openshift_min_t/openshift_min_app_t
14. What about trust between nodes.
● IPTables not enough
● Node1:Gear1 can not attack Node1:gear2
● Node1:Gear1 can attack Node2:gear2
● Labeled Networking between Nodes
– Based on MLS CIPSO Labels
● Labeled Networking SELinux rules
– Node1:Gear1 can use Node2:gear1
– Node1:Gear1 attacking Node2:gear2 blocked
● Requires UID being the same between nodes.
15. Problems with SELinux
Confinement
● Node Separation
– 127.0.0.1 blocked to all.
– We do not want multiple Domains binding to
127.0.0.1:8080
● First one wins
● Apps trying to do SELinux stuff
● SELinux blocks access to processes but it
knows they are there.
17. Containers != Security
● Running root in a container, machine pwned
● Local Privilege Escalation, machine pwned
● Much of the system is not containerized.
– Audit
– /sys
● selinuxfs, cgroupfs, sysfs
– Need to block mount
– Need to block mknod
18. Linux Namespaces
● Mount : mounting/unmounting filesystems
– Currently used by Openshift for /tmp, /var/tmp and /dev/shm
● UTS : hostname, domainname
● IPC : SysV message queues, semaphore/shared memory
segments
● Network: IPv4/IPv6 stacks, routing, firewall, proc/net
/sys/class/net directory trees, sock
– Critical to fix localhost problem
● Pid: Private /proc, multiple pid 1's
● UID: Just showing up in the Kernel now..
19. Libvirt-lxc
● Boot “init” binary
● SELinux Types + MCS
● Firewall ebtables/ip[6]tables
● Host FS passthrough bind mounts
● CGroups resource control
● Available in RHEL6.4
– But your on your own...
20. virt-sandbox
● Package to help managing Linux Containers
When writing SELinux policy, the first thing to understand, what is your security goal. For most people the security goal is to get to as close as minimal access to allow the confined application to get its job done and prevent its ability to effect other applications. For a lot of applications, you can configure the application to run in different Ways. Ftp for example can be configured to allow anonymous access to files, or access to users home directories, or access to the entire system. When you have an application like this, you can use booleans to allow administrators to reconfigure the policy, for their environment. When you are writing policy it is always good to ask experts about the policy you have written to see if you are allowing more access then necessary or if they know a better way to write the policy.
This slide shows one Virtual machine running as svirt_t:MCS1 and the other virtual machine running as svirt_t:MCS2. Which their image files labeled as svirt_image_t:MCS1 and svirt_image_t:MCS2. The same attack we saw before is being blocked by SELinux in the host kernel, and this protects Host as well as all virtual machines from attacking each other.